Towards End-to-End Encryption for Direct Messages in the Fediverse

Towards End-to-End Encryption for Direct Messages in the FediverseSkip to the content

As Twitter’s new management continues to nosedive the platform directly into the ground, many people are migrating to what seem like drop-in alternatives; i.e. Cohost and Mastodon. Some are even considering new platforms that none of us have heard of before (one is called “Hive”).

Needless to say, these are somewhat chaotic times.

One topic that has come up several times in the past few days, to the astonishment of many new Mastodon users, is that Direct Messages between users aren’t end-to-end encrypted.

And while that fact makes Mastodon DMs no less safe than Twitter DMs have been this whole time, there is clearly a lot of value and demand in deploying end-to-end encryption for ActivityPub (the protocol that Mastodon and other Fediverse software uses to communicate).

However, given that Melon Husk apparently wants to hurriedly ship end-to-end encryption (E2EE) in Twitter, in some vain attempt to compete with Signal, I took it upon myself to kickstart the E2EE effort for the Fediverse.

Archived

So I’d like to share my thoughts about E2EE, how to design such a system from the ground up, and why the direction Twitter is heading looks to be security theater rather than serious cryptographic engineering.

If you’re not interested in those things, but are interested in what I’m proposing for the Fediverse, head on over to the GitHub repository hosting my work-in-progress proposal draft as I continue to develop it.

How to Quickly Build E2EE

If one were feeling particularly cavalier about your E2EE designs, they could just generate then dump public keys through a server they control, pass between users, and have them encrypt client-side. Over and done. Check that box.

Every public key would be ephemeral and implicitly trusted, and the threat model would mostly be, “I don’t want to deal with law enforcement data requests.”

Hell, I’ve previously written an incremental blog post to teach developers about E2EE that begins with this sort of design. Encrypt first, ratchet second, manage trust relationships on public keys last.

If you’re catering to a slightly tech-savvy audience, you might throw in SHA256(pk1 + pk2) -> hex2dec() and call it a fingerprint / safety number / “conversation key” and not think further about this problem.

Look, technical users can verify out-of-band that they’re not being machine-in-the-middle attacked by our service.

An absolute fool who thinks most people will ever do this

From what I’ve gathered, this appears to be the direction that Twitter is going.

Archived

Now, if you’re building E2EE into a small hobby app that you developed for fun (say: a World of Warcraft addon for erotic roleplay chat), this is probably good enough.

If you’re building a private messaging feature that is intended to “superset Signal” for hundreds of millions of people, this is woefully inadequate.

Archived

If this is, indeed, the direction Musk is pushing what’s left of Twitter’s engineering staff, here is a brief list of problems with what they’re doing.

1. Twitter Web. How do you access your E2EE DMs after opening Twitter in your web browser on a desktop computer?
   • If you can, how do you know twitter.com isn’t including malicious JavaScript to snarf up your secret keys on behalf of law enforcement or a nation state with a poor human rights record?
   • If you can, how are secret keys managed across devices?
   • If you use a password to derive a secret key, how do you prevent weak, guessable, or reused passwords from weakening the security of the users’ keys?
   • If you cannot, how do users decide which is their primary device? What if that device gets lost, stolen, or damaged?
2. Authenticity. How do you reason about the person you’re talking with?
3. Forward Secrecy. If your secret key is compromised today, can you recover from this situation? How will your conversation participants reason about your new Conversation Key?
4. Multi-Party E2EE. If a user wants to have a three-way E2EE DM with the other members of their long-distance polycule, does Twitter enable that?
   • How are media files encrypted in a group setting? If you fuck this up, you end up like Threema.
   • Is your group key agreement protocol vulnerable to insider attacks?
5. Cryptography Implementations.
   • What does the KEM look like? If you’re using ECC, which curve? Is a common library being used in all devices?
   • How are you deriving keys? Are you just using the result of an elliptic curve (scalar x point) multiplication directly without hashing first?
6. Independent Third-Party Review.
   • Who is reviewing your protocol designs?
   • Who is reviewing your cryptographic primitives?
   • Who is reviewing the code that interacts with E2EE?
   • Is there even a penetration test before the feature launches?

As more details about Twitter’s approach to E2EE DMs come out, I’m sure the above list will be expanded with even more questions and concerns.

My hunch is that they’ll reuse liblithium (which uses Curve25519 and Gimli) for Twitter DMs, since the only expert I’m aware of in Musk’s employ is the engineer that developed that library for Tesla Motors. Whether they’ll port it to JavaScript or just compile to WebAssembly is hard to say.

How To Safely Build E2EE

You first need to decompose the E2EE problem into five separate but interconnected problems.

1. Client-Side Secret Key Management.
   • Multi-device support
   • Protect the secret key from being pilfered (i.e. by in-browser JavaScript delivered from the server)
2. Public Key Infrastructure and Trust Models.
   • TOFU (the SSH model)
   • X.509 Certificate Authorities
   • Certificate/Key/etc. Transparency
   • SigStore
   • PGP’s Web Of Trust
3. Key Agreement.
   • While this is important for 1:1 conversations, it gets combinatorially complex when you start supporting group conversations.
4. On-the-Wire Encryption.
   • Direct Messages
   • Media Attachments
   • Abuse-resistance (i.e. message franking for abuse reporting)
5. The Construction of the Previous Four.
   • The vulnerability of most cryptographic protocols exists in the joinery between the pieces, not the pieces themselves. For example, Matrix.

This might not be obvious to someone who isn’t a cryptography engineer, but each of those five problems is still really hard.

To wit: The latest IETF RFC draft for Message Layer Security, which tackles the Key Agreement problem above, clocks in at 137 pages.

Additionally, the order I specified these problems matters; it represents my opinion of which problem is relatively harder than the others.

When Twitter’s CISO, Lea Kissner, resigned, they lost a cryptography expert who was keenly aware of the relative difficulty of the first problem.

Archived

You may also notice the order largely mirrors my previous guide on the subject, in reverse. This is because teaching a subject, you start with the simplest and most familiar component. When you’re solving problems, you generally want the opposite: Solve the hardest problems first, then work towards the easier ones.

This is precisely what I’m doing with my E2EE proposal for the Fediverse.

The Journey of a Thousand Miles Begins With A First Step

Before you write any code, you need specifications.

Before you write any specifications, you need a threat model.

Before you write any threat models, you need both a clear mental model of the system you’re working with and how the pieces interact, and a list of security goals you want to achieve.

Less obviously, you need a specific list of non-goals for your design: Properties that you will not prioritize. A lot of security engineering involves trade-offs. For example: elliptic curve choice for digital signatures is largely a trade-off between speed, theoretical security, and real-world implementation security.

If you do not clearly specify your non-goals, they still exist implicitly. However, you may find yourself contradicting them as you change your mind over the course of development.

Being wishy-washy about your security goals is a good way to compromise the security of your overall design.

In my Mastodon E2EE proposal document, I have a section called Design Tenets, which states the priorities used to make trade-off decisions. I chose Usability as the highest priority, because of AviD’s Rule of Usability.

Security at the expense of usability comes at the expense of security.

Avi Douglen, Security StackExchange

Underneath Tenets, I wrote Anti-Tenets. These are things I explicitly and emphatically do not want to prioritize. Interoperability with any incumbent designs (OpenPGP, Matrix, etc.) is the most important anti-tenet when it comes to making decisions. If our end-state happens to interop with someone else’s design, cool. I’m not striving for it though!

Finally, this section concludes with a more formal list of Security Goals for the whole project.

Every component (from the above list of five) in my design will have an additional dedicated Security Goals section and Threat Model. For example: Client-Side Secret Key Management.

You will then need to tackle each component independently. The threat model for secret-key management is probably the trickiest. The actual encryption of plaintext messages and media attachments is comparatively simple.

Finally, once all of the pieces are laid out, you have the monumental (dare I say, mammoth) task of stitching them together into a coherent, meaningful design.

If you did your job well at the outset, and correctly understand the architecture of the distributed system you’re working with, this will mostly be straightforward.

Making Progress

At every step of the way, you do need to stop and ask yourself, “If I was an absolute chaos gremlin, how could I fuck with this piece of my design?” The more pieces your design has, the longer the list of ways to attack it will grow.

It’s also helpful to occasionally consider formal methods and security proofs. This can have surprising implications for how you use some algorithms.

You should also be familiar enough with the cryptographic primitives you’re working with before you begin such a journey; because even once you’ve solved the key management story (problems 1, 2 and 3 from the above list of 5), cryptographic expertise is still necessary.

• If you’re feeding data into a hash function, you should also be thinking about domain separation. More information.
• If you’re feeding data into a MAC or signature algorithm, you should also be thinking about canonicalization attacks. More information.
• If you’re encrypting data, you should be thinking about multi-key attacks and confused deputy attacks. Also, the cryptographic doom principle if you’re not using IND-CCA3 algorithms.
• At a higher-level, you should proactively defend against algorithm confusion attacks.

How Do You Measure Success?

It’s tempting to call the project “done” once you’ve completed your specifications and built a prototype, and maybe even published a formal proof of your design, but you should first collect data on every important metric:

1. How easy is it to use your solution?
2. How hard is it to misuse your solution?
3. How easy is it to attack your solution? Which attackers have the highest advantage?
4. How stable is your solution?
5. How performant is your solution? Are the slow pieces the deliberate result of a trade-off? How do you know the balance was struck corectly?

Where We Stand Today

I’ve only begun writing my proposal, and I don’t expect it to be truly ready for cryptographers or security experts to review until early 2023.

However, my clearly specified tenets and anti-tenets were already useful in discussing my proposal on the Fediverse.

@soatok @fasterthanlime Should probably embed the algo used for encryption in the data used for storing the encrypted blob, to support multiples and future changes.

@[email protected] proposes in-band protocol negotiation instead of versioned protocols

The main things I wanted to share today are:

1. The direction Twitter appears to be heading with their E2EE work, and why I think it’s a flawed approach
2. Designing E2EE requires a great deal of time, care, and expertise; getting to market quicker at the expense of a clear and careful design is almost never the right call

Mastodon? ActivityPub? Fediverse? OMGWTFBBQ!

In case anyone is confused about Mastodon vs ActivityPub vs Fediverse lingo:

The end goal of my proposal is that I want to be able to send DMs to queer furries that use Mastodon such that only my recipient can read them.

Achieving this end goal almost exclusively requires building for ActivityPub broadly, not Mastodon specifically.

However, I only want to be responsible for delivering this design into the software I use, not for every single possible platform that uses ActivityPub, nor all the programming languages they’re written in.

I am going to be aggressive about preventing scope creep, since I’m doing all this work for free. (I do have a Ko-Fi, but I won’t link to it from here. Send your donations to the people managing the Mastodon instance that hosts your account instead.)

My hope is that the design documents and technical specifications become clear enough that anyone can securely implement end-to-end encryption for the Fediverse–even if special attention needs to be given to the language-specific cryptographic libraries that you end up using.

Why Should We Trust You to Design E2EE?

This sort of question comes up inevitably, so I’d like to tackle it preemptively.

My answer to every question that begins with, “Why should I trust you” is the same: You shouldn’t.

There are certainly cryptography and cybersecurity experts that you will trust more than me. Ask them for their expert opinions of what I’m designing instead of blanketly trusting someone you don’t know.

I’m not interested in revealing my legal name, or my background with cryptography and computer security. Credentials shouldn’t matter here.

If my design is good, you should be able to trust it because it’s good, not because of who wrote it.

If my design is bad, then you should trust whoever proposes a better design instead. Part of why I’m developing it in the open is so that it may be forked by smarter engineers.

Knowing who I am, or what I’ve worked on before, shouldn’t enter your trust calculus at all. I’m a gay furry that works in the technology industry and this is what I’m proposing. Take it or leave it.

---

Credit for art used in header: LvJ, Harubaki